Claude Code cuts first-token wait for long conversations by 35%, reduces request data by 99.8%

Beating AI News Flash: Anthropic has launched Message Threads in public beta on the Messages API, reducing the overhead of AI Agents repeatedly uploading conversation history.

Previously, Coding Agents had to resend previous messages and tool execution results every time they called Claude. Now, the full conversation only needs to be uploaded once initially, and subsequent rounds only send new messages and the previous round's reply ID. The server temporarily stores the history, eliminating the time spent on repeated transmission, validation, and tokenization.

This feature is enabled by default in Claude Code. Anthropic disclosed that for requests of 500,000 to 1 million tokens, the median time to first token is reduced by about 35%. The median single-request size for long conversations has shrunk from about 800KB to 1.8KB, a reduction of approximately 99.8%.

However, the model still needs to process the full context, so token usage and costs remain unchanged. Message threads stored on the server also expire after 5 minutes of inactivity, at which point the history must be re-uploaded.